HyperFlow H3 Flash: Video Rebirth's 14B, 6-step MiniMax-H3 that makes video and sound on one consumer GPU

What it is, per the README. A few-step distilled version of MiniMax-H3 that generates 24 fps video (1344x768 by default) and 32 kHz stereo audio together. Three tasks: text to video+audio, first frame (or first and last frame) to video+audio, and a reference mode that takes up to 12 inputs (images in GGUF; images, video and audio in the INT8 build). Default is 6 steps; 4 for speed, 8 for quality.

Speed claims (Video Rebirth's numbers for a 5.2 s clip at 6 steps): 8.7 s on 4x B200, 36 s on one H200 and 92 s on one RTX 5090. NVIDIA's SOL-H3 engine runs it in real time on four Blackwell GPUs. The Q4_K_M GGUF is 9.1 GB and is measured on 24 and 32 GB cards in PCs with 16 GB of system RAM; 16 GB GPUs are listed as not yet measured.

How to run it. In ComfyUI, add ComfyUI-GGUF plus Video Rebirth's ComfyUI-HyperFlow-Flash nodes, download the 15.7 GB Qwen3-VL text encoder, and open the hyperflow_t2v, i2v, flf2v or ref2v templates. On 24 GB+ cards the new single-file INT8 checkpoints skip GGUF entirely; Video Rebirth says they run 5 to 15% faster than Q4_K_M on an RTX 5090 or 4090. There is also a Python CLI and a diffusers pipeline, plus stable-diffusion.cpp and Unsloth Studio paths.

Licence, read before downloading. The code is Apache-2.0, but the weights are a MiniMax-H3 derivative under the MiniMax H3 Community License: not licensed in the EU, UK, South Korea or the US without MiniMax's prior authorization, plus an acceptable use policy. Agencies in those regions should apply through MiniMax first.

Signals (HOT). Official launch on Nov 6 (about 391 likes), then Video Rebirth's own download update took off (about 696 likes for '15,000 downloads in under 24 hours'; they reported 40,000+ within 48 hours, and Hugging Face showed about 64k downloads when we checked). A Show HN reached 82 points, and the model is on Hugging Face's trending list. We missed it on Nov 6, so this is late by our own standard. We have not run it ourselves.

Our take: this is the cheapest open way so far to get H3-class video with matching sound on a single gaming card, and it pairs with the H3 LoRAs and ComfyUI tools we have already listed. Start with a 5-second Q4_K_M or INT8 test on a 24 GB card, and sort out the licence question before using it for client work in the US.